U.S. patent application number 17/302442 was filed with the patent office on 2021-12-30 for utilizing multiple stacked machine learning models to detect deepfake content.
The applicant listed for this patent is Accenture Global Solutions Limited. Invention is credited to Leah DING, Neil Hayden LIBERMAN.
Application Number | 20210406568 17/302442 |
Document ID | / |
Family ID | 1000005610749 |
Filed Date | 2021-12-30 |
United States Patent
Application |
20210406568 |
Kind Code |
A1 |
LIBERMAN; Neil Hayden ; et
al. |
December 30, 2021 |
UTILIZING MULTIPLE STACKED MACHINE LEARNING MODELS TO DETECT
DEEPFAKE CONTENT
Abstract
A device may receive training data for training a first machine
learning model, a second machine learning model, and a third
machine learning model and may train the first machine learning
model, the second machine learning model, and the third machine
learning model with the training data. The device may receive input
content and may process the input content, with the first machine
learning model, the second machine learning model, and the third
machine learning model, to generate a first model result, a second
model result, and a third model result, respectively. The device
may process the first model result, the second model result, and
the third model result, with an aggregation model, to generate a
detection result indicating whether the particular input content is
a deepfake or is real and may perform one or more actions based on
the detection result.
Inventors: |
LIBERMAN; Neil Hayden;
(Arlington, VA) ; DING; Leah; (North Potomac,
MD) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Accenture Global Solutions Limited |
Dublin |
|
IE |
|
|
Family ID: |
1000005610749 |
Appl. No.: |
17/302442 |
Filed: |
May 3, 2021 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
63043392 |
Jun 24, 2020 |
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06K 9/6257 20130101;
G06K 9/00281 20130101; G06K 9/00899 20130101; G06T 3/40 20130101;
G06T 2207/30201 20130101; G06Q 20/3821 20130101; G06T 7/11
20170101; G06Q 50/265 20130101; G06N 3/08 20130101; G06T 2207/20084
20130101; G06T 2207/20132 20130101; G06T 2207/20081 20130101; G06N
3/0454 20130101 |
International
Class: |
G06K 9/00 20060101
G06K009/00; G06N 3/04 20060101 G06N003/04; G06N 3/08 20060101
G06N003/08; G06Q 20/38 20060101 G06Q020/38; G06Q 50/26 20060101
G06Q050/26; G06T 7/11 20060101 G06T007/11; G06T 3/40 20060101
G06T003/40; G06K 9/62 20060101 G06K009/62 |
Claims
1. A method, comprising: receiving, by a device, training data that
includes input content for training machine learning models that
include a first machine learning model, a second machine learning
model, and a third machine learning model; preprocessing, by the
device, the training data to generate preprocessed training data;
training, by the device, the machine learning models, with the
preprocessed training data, to generate trained machine learning
models that include a first trained machine learning model, a
second trained machine learning model, and a third trained machine
learning model; receiving, by the device, particular input content;
processing, by the device, the particular input content, with the
first trained machine learning model, to generate a first model
result; processing, by the device, the particular input content,
with the second trained machine learning model, to generate a
second model result; processing, by the device, the particular
input content, with the third trained machine learning model, to
generate a third model result; processing, by the device, the first
model result, the second model result, and the third model result,
with an aggregation model, to generate a detection result
indicating whether the particular input content is a deepfake or is
real; and performing, by the device, one or more actions based on
the detection result.
2. The method of claim 1, wherein preprocessing the training data
to generate the preprocessed training data comprises: extracting
frames from the input content; extracting images of faces from the
frames; returning facial landmarks, associated with the images of
the faces, to the frames; and adding perturbations to the frames,
after returning the facial landmarks, to generate the preprocessed
training data.
3. The method of claim 2, wherein adding the perturbations to the
frames comprises: cropping outer pixels of the images of the faces;
and replacing the cropped outer pixels with a border or resizing
the images of the faces based on the cropped outer pixels.
4. The method of claim 1, wherein each of the first model result,
the second model result, and the third model result provides an
indication of whether the particular input content is a deepfake or
is real.
5. The method of claim 1, wherein the first machine learning model
is a full-face image convolutional neural network (CNN) model,
wherein the second machine learning model is a partial image CNN
model, and wherein the third machine learning model is a support
vector machine learning model.
6. The method of claim 1, wherein the aggregation model is a
logistic regression machine learning model.
7. The method of claim 1, wherein the first model result is a first
probability that the particular input content is a deepfake and the
second model result is a second probability that the particular
input content is a deepfake.
8. A device, comprising: one or more memories; and one or more
processors, coupled to the one or more memories, configured to:
receive input content; process the input content, with a first
machine learning model, to generate a first model result; process
the input content, with a second machine learning model, to
generate a second model result; process the input content, with a
third machine learning model, to generate a third model result;
process the first model result, the second model result, and the
third model result, with an aggregation model, to generate a
detection result indicating whether the input content is a deepfake
or is real; and perform one or more actions based on the detection
result.
9. The device of claim 8, wherein the one or more processors, to
process the input content, with the third machine learning model,
to generate the third model result, are configured to: process the
input content, with the third machine learning model, to generate a
plurality of binary predictions that the input content is a
deepfake; and average the plurality of binary predictions to
determine a probability that the input content is a deepfake,
wherein the third model result corresponds to the probability.
10. The device of claim 8, wherein the one or more processors, to
perform the one or more actions, are configured to one or more of:
provide the detection result for display; or retrain one or more of
the first machine learning model, the second machine learning
model, or the third machine learning model based on the detection
result.
11. The device of claim 8, wherein the one or more processors, to
perform the one or more actions, are configured to one or more of:
generate and provide a notification to a law enforcement agency
when the detection result indicates that the input content is a
deepfake; or generate and provide a notification to an entity
associated with the input content when the detection result
indicates that the input content is a deepfake.
12. The device of claim 8, wherein the one or more processors, to
perform the one or more actions, are configured to: correct the
input content to remove deepfake content and to generate modified
input content; and utilize the modified input content.
13. The device of claim 8, wherein the one or more processors, to
perform the one or more actions, are configured to: prevent a
financial transaction when the detection result indicates that the
input content is a deepfake of a credential required for performing
the financial transaction.
14. The device of claim 8, wherein the detection result is a binary
result with a first value indicating that the input content is a
deepfake and a second value indicating that the input content is
real.
15. A non-transitory computer-readable medium storing a set of
instructions, the set of instructions comprising: one or more
instructions that, when executed by one or more processors of a
device, cause the device to: receive input content; process the
input content, with a first machine learning model, to generate a
first model result; process the input content, with a second
machine learning model, to generate a second model result; process
the input content, with a third machine learning model, to generate
a third model result, wherein the first machine learning model, the
second machine learning model, and the third machine learning are
trained based on training data that includes preprocessed
historical input content; process the first model result, the
second model result, and the third model result, with an
aggregation model, to generate a detection result indicating
whether the input content is a deepfake or is real; and perform one
or more actions based on the detection result.
16. The non-transitory computer-readable medium of claim 15,
wherein the preprocessed historical input content includes
perturbations.
17. The non-transitory computer-readable medium of claim 15,
wherein the first model result is a first probability that the
input content is a deepfake and the second model result is a second
probability that the input content is a deepfake.
18. The non-transitory computer-readable medium of claim 15,
wherein the one or more instructions, that cause the device to
process the input content, with the third machine learning model,
to generate the third model result, cause the device to: process
the input content, with the third machine learning model, to
generate a plurality of binary predictions that the input content
is a deepfake; and average the plurality of binary predictions to
determine a probability that the input content is a deepfake,
wherein the third model result corresponds to the probability.
19. The non-transitory computer-readable medium of claim 15,
wherein the one or more instructions, that cause the device to
perform the one or more actions, cause the device to one or more
of: provide the detection result for display; retrain one or more
of the machine learning models based on the detection result;
generate and provide a notification to a law enforcement agency
when the detection result indicates that the input content is a
deepfake; or generate and provide a notification to an entity
associated with the input content when the detection result
indicates that the input content is a deepfake.
20. The non-transitory computer-readable medium of claim 15,
wherein the one or more instructions, that cause the device to
perform the one or more actions, cause the device to: correct the
input content to remove deepfake content and to generate modified
input content; and utilize the modified input content.
Description
CROSS-REFERENCE TO RELATED APPLICATION
[0001] This patent application claims priority to Provisional
Patent Application No. 63/043,392, filed on Jun. 24, 2020, and
entitled "UTILIZING MULTIPLE STACKED MACHINE LEARNING MODELS TO
DETECT DEEPFAKE CONTENT." The disclosure of the prior application
is considered part of and is incorporated by reference into this
patent application.
BACKGROUND
[0002] Digital content (e.g., images, audio, video, and/or the
like) can be manipulated with artificial intelligence-based
solutions. For example, an image (e.g., a company's design
sketches) can be manipulated by artificial intelligence (e.g., a
deepfake model) to generate substantially similar images (e.g.,
designs) that infringe the company's design sketches. Audio can be
manipulated by artificial intelligence for voice impersonation.
Moreover, images and videos can be combined and superimposed onto
source images or videos using artificial intelligence.
SUMMARY
[0003] In some implementations, a method may include receiving
training data that includes input content for training machine
learning models that include a first machine learning model, a
second machine learning model, and a third machine learning model
and preprocessing the training data to generate preprocessed
training data. The method may include training the machine learning
models, with the preprocessed training data, to generate trained
machine learning models that include a first trained machine
learning model, a second trained machine learning model, and a
third trained machine learning model and receiving particular input
content. The method may include processing the particular input
content, with the first trained machine learning model, to generate
a first model result, and processing the particular input content,
with the second trained machine learning model, to generate a
second model result. The method may include processing the
particular input content, with the third trained machine learning
model, to generate a third model result, and processing the first
model result, the second model result, and the third model result,
with an aggregation model, to generate a detection result
indicating whether the particular input content is a deepfake or is
real. The method may include performing one or more actions based
on the detection result.
[0004] In some implementations, a device includes one or more
memories and one or more processors to receive input content and
process the input content, with a first machine learning model, to
generate a first model result. The one or more processors may
process the input content, with a second machine learning model, to
generate a second model result, and may process the input content,
with a third machine learning model, to generate a third model
result. The one or more processors may process the first model
result, the second model result, and the third model result, with
an aggregation model, to generate a detection result indicating
whether the input content is a deepfake or is real and may perform
one or more actions based on the detection result.
[0005] In some implementations, a non-transitory computer-readable
medium may store a set of instructions that includes one or more
instructions that, when executed by one or more processors of a
device, cause the device to receive input content and process the
input content, with a first machine learning model, to generate a
first model result. The one or more instructions may cause the
device to process the input content, with a second machine learning
model, to generate a second model result and process the input
content, with a third machine learning model, to generate a third
model result, wherein the first machine learning model, the second
machine learning model, and the third machine learning are trained
based on training data that includes preprocessed historical input
content. The one or more instructions may cause the device to
process the first model result, the second model result, and the
third model result, with an aggregation model, to generate a
detection result indicating whether the input content is a deepfake
or is real and perform one or more actions based on the detection
result.
BRIEF DESCRIPTION OF THE DRAWINGS
[0006] FIGS. 1A-1F are diagrams of an example implementation
described herein.
[0007] FIG. 2 is a diagram illustrating an example of training and
using a machine learning model in connection with detecting
deepfake content.
[0008] FIG. 3 is a diagram of an example environment in which
systems and/or methods described herein may be implemented.
[0009] FIG. 4 is a diagram of example components of one or more
devices of FIG. 3.
[0010] FIG. 5 is a flowchart of an example process for utilizing
multiple stacked machine learning models to detect deepfake
content.
DETAILED DESCRIPTION
[0011] The following detailed description of example
implementations refers to the accompanying drawings. The same
reference numbers in different drawings may identify the same or
similar elements.
[0012] Digital content manipulation with artificial intelligence
has been used to create fake celebrity videos, fake news, malicious
hoaxes, and/or the like. A deepfake model is an artificial
intelligence-based technology (e.g., deep learning technology) used
to produce or alter digital content so that the digital content
presents something that didn't, in fact, occur. Deepfake detection
is the task of identifying manipulated faces in video or image
format. These manipulations can be as subtle as a change to a
facial expression or as drastic as a person's face being swapped
with another face. While there have been models showing progress in
this area, a major issue of deepfake detection is that of
generalizability. It is a relatively simple task to achieve good
accuracy on a set of training videos. However, it can be very
difficult to achieve similar accuracy on test data. Furthermore,
the task becomes increasingly difficult when analyzing forgeries
generated from models that are not used in the training set.
[0013] Therefore, current deepfake detection techniques consume
computing resources (e.g., processing resources, memory resources,
communication resources, and/or the like), networking resources,
and/or the like associated with providing less accurate deepfake
determinations for digital content, correcting deepfake
determinations for digital content, handling theft of proprietary
information based on deepfake credentials, and/or the like.
[0014] Some implementations described herein relate to a detection
system that utilizes multiple stacked machine learning models to
detect deepfake content. For example, the detection system may
receive training data that includes input content for training
machine learning models that include a first machine learning
model, a second machine learning model, and a third machine
learning model and may preprocess the training data to generate
preprocessed training data. The method may include training the
machine learning models, with the preprocessed training data, to
generate trained machine learning models that include a first
trained machine learning model, a second trained machine learning
model, and a third trained machine learning model and receiving
particular input content. The method may include processing the
particular input content, with the first trained machine learning
model, to generate a first model result, and processing the
particular input content, with the second trained machine learning
model, to generate a second model result. The method may include
processing the particular input content, with the third trained
machine learning model, to generate a third model result, and
processing the first model result, the second model result, and the
third model result, with an aggregation model, to generate a
detection result indicating whether the particular input content is
a deepfake or is real. The method may include performing one or
more actions based on the detection result.
[0015] In this way, the detection system utilizes multiple stacked
machine learning models to detect deepfake content. The detection
system may utilize numerous machine learning models that provide
different detection methods that are then aggregated into a single
model that outputs a binary classification of content as real or
fake. Additionally, the detection system utilizes added
perturbations to content which may lead to generalization
improvements. This, in turn, conserves computing resources,
networking resources, and/or the like that would otherwise have
been consumed in providing less accurate deepfake determinations
for digital content, correcting deepfake determinations for digital
content, handling theft of proprietary information based on
deepfake credentials, and/or the like.
[0016] FIGS. 1A-1F are diagrams of an example 100 associated with
utilizing multiple stacked machine learning models to detect
deepfake content. As shown in FIGS. 1A-1F, example 100 includes a
user device and a detection system. The user device may include a
desktop computer, a mobile telephone, a virtual reality device,
wearable communication device, a laptop computer, and/or the like
associated with a user. The detection system may include a system
that utilizes multiple stacked machine learning models to detect
deepfake content. Further details of the user device and the
detection system are provided elsewhere herein.
[0017] As shown in FIG. 1A, and by reference number 105, the
detection system may receive training data that includes input
content (e.g., images, videos, audio, and/or the like) for training
machine learning models. For example, the detection system may
receive the training data from the user device or from one or more
other devices. The training data may include a first group of
historical content that has been manipulated via a deepfake model,
a second group of historical content that has not been manipulated
via a deepfake model, and/or the like. For example, the training
data may include a first group of historical videos where some or
all of the frames of the first group have been manipulated via a
deepfake model, a second group of historical videos where some or
all of the frames of the second group have been manipulated via a
deepfake model, and/or the like.
[0018] As further shown in FIG. 1A, the detection system may
preprocess the input content of the training data so that the
detection system may analyze each frame of the input content before
aggregating image results into a single deepfake prediction. For
example, as shown by reference number 110, the detection system may
extract frames from the input content. If the input content
includes videos, each video may include multiple frames (e.g.,
images) that, together, constitute each video. Thus, the detection
system may extract the multiple frames from each video to generate
multiple frames (e.g., multiple images).
[0019] As further shown in FIG. 1A, and by reference number 115,
the detection system may extract face images from the frames of the
input content. Because a deepfake may be designed to alter an image
of a face, the detection system may utilize the extracted face
images from frames rather than the entire frames. In some
implementations, a class breakdown of the training data may not be
inherently balanced (e.g., with more deepfake videos than real
unmanipulated videos), so the detection system may artificially
balance the training data by utilizing more frames from the real
videos than frames from the deepfake videos. After extracting the
face images from the frames of the input content, the detection
system may resize the face images to input dimensions necessary for
the machine learning models.
[0020] As further shown in FIG. 1A, and by reference number 120,
the detection system may return facial landmarks to the frames. For
example, the detection system may return the facial landmarks to
the extracted face images. In some implementations, the detection
system may utilize a library of landmarks to return the facial
landmarks to the extracted face images. With the facial landmarks,
the detection system may locate chin areas of the face images and
may extract portions (e.g., a square areas) of the chin areas as
input for one or more of the machine learning models.
[0021] As further shown in FIG. 1A, and by reference number 125,
the detection system may add perturbations to the frames (e.g., to
the extracted face images and after returning the facial landmarks
to the extracted face images) to generate preprocessed training
data. Adding small perturbations to test data may lead to a
significant loss of performance for many deepfake models. For this
reason, the detection system may add the perturbations to the
training data. For example, the detection system may add first
perturbations to the training data by cropping outer pixels of the
extracted face images and replacing the outer pixels with a border
(e.g., a black colored border). The cropping of the outer pixels
may result in images that sometimes cut off minor portions of
faces. However, useful information is not lost by cropping the
outer pixels since manipulations of face images almost always occur
around noses, eyes, and/or mouths, which should be located
centrally in the face images for people facing a camera. The
detection system may add second perturbations to the training data
by cropping outer pixels of the extracted face images and
stretching the face images to a predetermined input size. This may
create slightly different face images from the original uncropped
face images but may lower qualities of the face images and may
increase a variability of the training data. In some
implementations, the detection system may add the first and second
perturbations to a random quantity (e.g., ten percent, twenty
percent, and/or the like) of the face images.
[0022] As shown in FIG. 1B, and by reference number 130, the
detection system may train the machine learning models, with the
preprocessed training data, to generate trained machine learning
models. For example, the detection system may train a first machine
learning model, a second machine learning model, and a third
machine learning model, with the preprocessed training data, to
generate a first trained machine learning model, a second trained
machine learning model, and a third trained machine learning model,
respectively. In some implementations, the detection system may
concurrently train the first machine learning model, the second
machine learning model, and the third machine learning model with
the preprocessed training data; may sequentially train the first
machine learning model, the second machine learning model, and the
third machine learning model with the preprocessed training data;
and/or the like. Prior to training the machine learning models, the
detection system may label the preprocessed training data. For
example, the detection system may label the extracted face images
from real videos as "real," may label the extracted face images
from fake videos as "fake," and/or the like.
[0023] The first machine learning model may include a convolutional
neural network (CNN) model, referred to as a full-face image CNN
model. The detection system may utilize the labeled training data
with the full-face CNN model to determine classifications for the
training data (e.g., real (or not a deepfake), a deepfake, and/or
the like).
[0024] The second machine learning model may include a CNN model,
referred to as a partial image CNN model. The partial image CNN
model may be a variation of the full-face image CNN model. Because
blending is a universal step in creating deepfakes, focusing on a
partial area (e.g., a chin of a face image) improves generalization
of the partial image CNN model. The chin may be utilized as the
partial area of focus since blending always occurs in the chin
area. Furthermore, a first type of deepfake is a face swap of two
people, where blending occurs throughout many different regions of
a face image. A second type of deepfake is an alteration of a
facial expression or mouth movement, which is a much more subtle
deepfake that may not encompass as much blending around the entire
face image as the face swap. Thus, blending is expected to occur
near a mouth or chin of a face image and the partial image CNN
model may detect such blending.
[0025] The third machine learning model may include a support
vector machine (SVM) learning model, referred to as a blur SVM
learning model. Many fake images contain regions of faces that are
sharp while other regions of the faces are blurred. The detection
system may evaluate a level of blur across different regions of a
face and may calculate a difference between neighboring regions.
The detection system may utilize a Laplacian variance to compute
levels of blur for different regions and may calculate the
differences between a current region and neighboring regions based
on the computed levels of blur. The detection system may utilize
the differences to compute a mean, a median, and a standard
deviation as features for the SVM learning model. The detection
system may also calculate differences between individual pixel
values (e.g., by comparing each pixel value to neighboring pixel
values) and may utilize the differences to calculate a mean, a
median, and a standard deviation as features for the SVM learning
model.
[0026] As shown in FIG. 1C, and by reference number 135, the
detection system may receive particular input content that is a
deepfake or is real (e.g., not a deepfake). For example, the
detection system may receive the particular input content from the
user device. The particular input content may include a video, one
or more images, audio data, and/or the like. In some
implementations, the detection system receives the particular input
content as part of a request for a determination of whether the
particular input content is a deepfake or is real.
[0027] As shown in FIG. 1D, and by reference number 140, the
detection system may process the particular input content, with the
first trained machine learning model, to generate a first model
result. The first model result may include an indication of whether
the particular input content is a deepfake or is real (e.g., not a
deepfake). In some implementations, the first model result includes
a probability that the particular input content is a deepfake.
[0028] As further shown in FIG. 1D, and by reference number 145,
the detection system may process the particular input content, with
the second trained machine learning model, to generate a second
model result. The second model result may include an indication of
whether the particular input content is a deepfake or is real
(e.g., not a deepfake). In some implementations, the second model
result includes a probability that the particular input content is
a deepfake.
[0029] As further shown in FIG. 1D, and by reference number 150,
the detection system may process the particular input content, with
the third trained machine learning model, to generate a third model
result. The third model result may include an indication of whether
the particular input content is a deepfake or is real (e.g., not a
deepfake). In some implementations, the third trained machine
learning model processes the particular input content to generate a
plurality of binary predictions that the input content is a
deepfake. The third trained machine learning model may average the
plurality of binary predictions to determine a probability that the
particular input content is a deepfake. The third model result may
correspond to the probability that the particular input content is
a deepfake.
[0030] In some implementations, the detection system may
concurrently process the particular input content with the first
trained machine learning model, the second trained machine learning
model, and the third trained machine learning model; may
sequentially process the particular input content with the first
trained machine learning model, the second trained machine learning
model, and the third trained machine learning model; and/or the
like.
[0031] As shown in FIG. 1E, and by reference number 155, the
detection system may process the first model result, the second
model result, and the third model result, with an aggregation
model, to generate a detection result indicating whether the
particular input content is a deepfake or is real (e.g., not a
deepfake). The aggregation model may include a logistic regression
machine learning model. In some implementations, the detection
result is a final probability that the particular input content is
a deepfake. The final probability may be determined based on a
first probability that the particular input content is a deepfake
(e.g., generated by the first trained machine learning model), a
second probability that the particular input content is a deepfake
(e.g., generated by the second trained machine learning model), and
a third probability that the particular input content is a deepfake
(e.g., generated by the third trained machine learning model). In
some implementations, the detection result is a binary result with
a first value indicating that the particular input content is a
deepfake and a second value indicating that the particular input
content is real.
[0032] As shown in FIG. 1F, and by reference number 160, the
detection system may perform one or more actions based on the
detection result. In some implementations, the one or more actions
include the detection system providing the detection result for
display. For example, the detection system may provide the
detection result to the user device and the user device may display
the detection result to a user of the user device. The user may
utilize the detection result to block the particular input content,
modify the particular input content, and/or take another
appropriate action with respect to the particular input content. In
this way, the detection system conserves computing resources,
networking resources, and/or the like that would otherwise have
been consumed in providing less accurate deepfake determinations
for the particular input content, correcting deepfake
determinations for the particular input content, and/or the
like.
[0033] In some implementations, the one or more actions include the
detection system generating and providing a notification to a law
enforcement agency when the detection result indicates that the
particular input content is a deepfake. For example, when the
detection result indicates that the particular input content is a
deepfake, the detection system may provide a notification to a
device associated with the law enforcement agency so that the law
enforcement agency may take appropriate action against a creator of
the particular input content. In this way, the detection system
conserves computing resources, networking resources, and/or the
like that would otherwise have been consumed in handling theft of
proprietary information based on deepfake credentials, handling
legal ramifications of the particular input content, and/or the
like.
[0034] In some implementations, the one or more actions include the
detection system generating and providing a notification to an
entity associated with the particular input content when the
detection result indicates that the particular input content is a
deepfake. For example, when the detection result indicates that the
particular input content is a deepfake, the detection system may
provide a notification to the entity associated with the particular
input content so that the entity may block the particular input
content, modify the particular input content, and/or take another
appropriate action with respect to the particular input content. In
this way, the detection system conserves computing resources,
networking resources, and/or the like that would otherwise have
been consumed in providing less accurate deepfake determinations
for the particular input content, correcting deepfake
determinations for the particular input content, and/or the
like.
[0035] In some implementations, the one or more actions include the
detection system correcting the particular input content to remove
deepfake content and to generate modified input content, and
utilizing the modified input content. For example, if the
particular input content is to be provided on a website, the
detection system may correct the particular input content to
generate the modified input content, and may provide the modified
input content to the website. In this way, the detection system
conserves computing resources, networking resources, and/or the
like that would otherwise have been consumed in providing less
accurate deepfake determinations for the particular input content,
correcting deepfake determinations for the particular input
content, and/or the like.
[0036] In some implementations, the one or more actions include the
detection system preventing a financial transaction when the
detection result indicates that the particular input content is a
deepfake of a credential required for performing the financial
transaction. For example, the particular input content may include
a face image that is to be provided to verify a user for
performance of the financial transaction. When the detection result
indicates that the face image is a deepfake, the detection system
may prevent the financial transaction from occurring and may notify
a law enforcement agency about the attempted financial fraud. In
this way, the detection system conserves computing resources,
networking resources, and/or the like that would otherwise have
been consumed in handling financial fraud based on deepfake
credentials, handling legal ramifications of the financial fraud,
and/or the like.
[0037] In some implementations, the one or more actions include the
detection system retraining one or more of the machine learning
models based on the detection result. The detection system may
utilize the detection result as additional training data for
retraining the one or more of the machine learning models, thereby
increasing the quantity of training data available for training the
one or more of the machine learning models. Accordingly, the
detection system may conserve computing resources associated with
identifying, obtaining, and/or generating historical data for
training the one or more of the machine learning models relative to
other systems for identifying, obtaining, and/or generating
historical data for training machine learning models.
[0038] In this way, the detection system utilizes multiple stacked
machine learning models to detect deepfake content. The detection
system may utilize numerous machine learning models that provide
different detection methods that are then aggregated into a single
model that outputs a binary classification of content as real or
fake. Additionally, the detection system utilizes added
perturbations to content which may lead to generalization
improvements. This, in turn, conserves computing resources,
networking resources, and/or the like that would otherwise have
been consumed in providing less accurate deepfake determinations
for digital content, correcting deepfake determinations for digital
content, handling theft of proprietary information based on
deepfake credentials, and/or the like.
[0039] As indicated above, FIGS. 1A-1F are provided as an example.
Other examples may differ from what is described with regard to
FIGS. 1A-1F. The number and arrangement of devices shown in FIGS.
1A-1F are provided as an example. In practice, there may be
additional devices, fewer devices, different devices, or
differently arranged devices than those shown in FIGS. 1A-1F.
Furthermore, two or more devices shown in FIGS. 1A-1F may be
implemented within a single device, or a single device shown in
FIGS. 1A-1F may be implemented as multiple, distributed devices.
Additionally, or alternatively, a set of devices (e.g., one or more
devices) shown in FIGS. 1A-1F may perform one or more functions
described as being performed by another set of devices shown in
FIGS. 1A-1F.
[0040] FIG. 2 is a diagram illustrating an example 200 of training
and using a machine learning model (e.g., the machine learning
models or the aggregation model) in connection with detecting
deepfake content. The machine learning model training and usage
described herein may be performed using a machine learning system.
The machine learning system may include or may be included in a
computing device, a server, a cloud computing environment, and/or
the like, such as the detection system described in more detail
elsewhere herein.
[0041] As shown by reference number 205, a machine learning model
may be trained using a set of observations. The set of observations
may be obtained from historical data, such as data gathered during
one or more processes described herein. In some implementations,
the machine learning system may receive the set of observations
(e.g., as input) from the detection system, as described elsewhere
herein.
[0042] As shown by reference number 210, the set of observations
includes a feature set. The feature set may include a set of
variables, and a variable may be referred to as a feature. A
specific observation may include a set of variable values (or
feature values) corresponding to the set of variables. In some
implementations, the machine learning system may determine
variables for a set of observations and/or variable values for a
specific observation based on input received from the detection
system. For example, the machine learning system may identify a
feature set (e.g., one or more features and/or feature values) by
extracting the feature set from structured data, by performing
natural language processing to extract the feature set from
unstructured data, by receiving input from an operator, and/or the
like.
[0043] As an example, a feature set for a set of observations may
include a first feature of a first model result, a second feature
of a second model result, a third feature of a third model result,
and so on. As shown, for a first observation, the first feature may
have a value of first model result 1, the second feature may have a
value of second model result 1, the third feature may have a value
of third model result 1, and so on. These features and feature
values are provided as examples and may differ in other
examples.
[0044] As shown by reference number 215, the set of observations
may be associated with a target variable. The target variable may
represent a variable having a numeric value, may represent a
variable having a numeric value that falls within a range of values
or has some discrete possible values, may represent a variable that
is selectable from one of multiple options (e.g., one of multiple
classes, classifications, labels, and/or the like), may represent a
variable having a Boolean value, and/or the like. A target variable
may be associated with a target variable value, and a target
variable value may be specific to an observation. In example 200,
the target variable is a detection result, which has a value of
real for the first observation.
[0045] The target variable may represent a value that a machine
learning model is being trained to predict, and the feature set may
represent the variables that are input to a trained machine
learning model to predict a value for the target variable. The set
of observations may include target variable values so that the
machine learning model can be trained to recognize patterns in the
feature set that lead to a target variable value. A machine
learning model that is trained to predict a target variable value
may be referred to as a supervised learning model.
[0046] In some implementations, the machine learning model may be
trained on a set of observations that do not include a target
variable. This may be referred to as an unsupervised learning
model. In this case, the machine learning model may learn patterns
from the set of observations without labeling or supervision, and
may provide output that indicates such patterns, such as by using
clustering and/or association to identify related groups of items
within the set of observations.
[0047] As shown by reference number 220, the machine learning
system may train a machine learning model using the set of
observations and using one or more machine learning algorithms,
such as a regression algorithm, a decision tree algorithm, a neural
network algorithm, a k-nearest neighbor algorithm, a support vector
machine algorithm, and/or the like. After training, the machine
learning system may store the machine learning model as a trained
machine learning model 225 to be used to analyze new
observations.
[0048] As shown by reference number 230, the machine learning
system may apply the trained machine learning model 225 to a new
observation, such as by receiving a new observation and inputting
the new observation to the trained machine learning model 225. As
shown, the new observation may include a first feature of first
model result X, a second feature of second model result Y, a third
feature of third model result Z, and so on, as an example. The
machine learning system may apply the trained machine learning
model 225 to the new observation to generate an output (e.g., a
result). The type of output may depend on the type of machine
learning model and/or the type of machine learning task being
performed. For example, the output may include a predicted value of
a target variable, such as when supervised learning is employed.
Additionally, or alternatively, the output may include information
that identifies a cluster to which the new observation belongs,
information that indicates a degree of similarity between the new
observation and one or more other observations, and/or the like,
such as when unsupervised learning is employed.
[0049] As an example, the trained machine learning model 225 may
predict a value of fake for the target variable of the detection
result for the new observation, as shown by reference number 235.
Based on this prediction, the machine learning system may provide a
first recommendation, may provide output for determination of a
first recommendation, may perform a first automated action, may
cause a first automated action to be performed (e.g., by
instructing another device to perform the automated action), and/or
the like.
[0050] In some implementations, the trained machine learning model
225 may classify (e.g., cluster) the new observation in a cluster,
as shown by reference number 240. The observations within a cluster
may have a threshold degree of similarity. As an example, if the
machine learning system classifies the new observation in a first
cluster (e.g., a first model result cluster), then the machine
learning system may provide a first recommendation. Additionally,
or alternatively, the machine learning system may perform a first
automated action and/or may cause a first automated action to be
performed (e.g., by instructing another device to perform the
automated action) based on classifying the new observation in the
first cluster.
[0051] As another example, if the machine learning system were to
classify the new observation in a second cluster (e.g., a second
model result cluster), then the machine learning system may provide
a second (e.g., different) recommendation and/or may perform or
cause performance of a second (e.g., different) automated
action.
[0052] In some implementations, the recommendation and/or the
automated action associated with the new observation may be based
on a target variable value having a particular label (e.g.,
classification, categorization, and/or the like), may be based on
whether a target variable value satisfies one or more thresholds
(e.g., whether the target variable value is greater than a
threshold, is less than a threshold, is equal to a threshold, falls
within a range of threshold values, and/or the like), may be based
on a cluster in which the new observation is classified, and/or the
like.
[0053] In this way, the machine learning system may apply a
rigorous and automated process to detect deepfake content. The
machine learning system enables recognition and/or identification
of tens, hundreds, thousands, or millions of features and/or
feature values for tens, hundreds, thousands, or millions of
observations, thereby increasing accuracy and consistency and
reducing delay associated with detecting deepfake content relative
to requiring computing resources to be allocated for tens,
hundreds, or thousands of operators to manually detect deepfake
content.
[0054] As indicated above, FIG. 2 is provided as an example. Other
examples may differ from what is described in connection with FIG.
2.
[0055] FIG. 3 is a diagram of an example environment 300 in which
systems and/or methods described herein may be implemented. As
shown in FIG. 3, environment 300 may include a detection system
301, which may include one or more elements of and/or may execute
within a cloud computing system 302. The cloud computing system 302
may include one or more elements 303-313, as described in more
detail below. As further shown in FIG. 3, environment 300 may
include a network 320 and/or a user device 330. Devices and/or
elements of environment 300 may interconnect via wired connections
and/or wireless connections.
[0056] The cloud computing system 302 includes computing hardware
303, a resource management component 304, a host operating system
(OS) 305, and/or one or more virtual computing systems 306. The
resource management component 304 may perform virtualization (e.g.,
abstraction) of computing hardware 303 to create the one or more
virtual computing systems 306. Using virtualization, the resource
management component 304 enables a single computing device (e.g., a
computer, a server, and/or the like) to operate like multiple
computing devices, such as by creating multiple isolated virtual
computing systems 306 from computing hardware 303 of the single
computing device. In this way, computing hardware 303 can operate
more efficiently, with lower power consumption, higher reliability,
higher availability, higher utilization, greater flexibility, and
lower cost than using separate computing devices.
[0057] Computing hardware 303 includes hardware and corresponding
resources from one or more computing devices. For example,
computing hardware 303 may include hardware from a single computing
device (e.g., a single server) or from multiple computing devices
(e.g., multiple servers), such as multiple computing devices in one
or more data centers. As shown, computing hardware 303 may include
one or more processors 307, one or more memories 308, one or more
storage components 309, and/or one or more networking components
310. Examples of a processor, a memory, a storage component, and a
networking component (e.g., a communication component) are
described elsewhere herein.
[0058] The resource management component 304 includes a
virtualization application (e.g., executing on hardware, such as
computing hardware 303) capable of virtualizing computing hardware
303 to start, stop, and/or manage one or more virtual computing
systems 306. For example, the resource management component 304 may
include a hypervisor (e.g., a bare-metal or Type 1 hypervisor, a
hosted or Type 2 hypervisor, and/or the like) or a virtual machine
monitor, such as when the virtual computing systems 306 are virtual
machines 311. Additionally, or alternatively, the resource
management component 304 may include a container manager, such as
when the virtual computing systems 306 are containers 312. In some
implementations, the resource management component 304 executes
within and/or in coordination with a host operating system 305.
[0059] A virtual computing system 306 includes a virtual
environment that enables cloud-based execution of operations and/or
processes described herein using computing hardware 303. As shown,
a virtual computing system 306 may include a virtual machine 311, a
container 312, a hybrid environment 313 that includes a virtual
machine and a container, and/or the like. A virtual computing
system 306 may execute one or more applications using a file system
that includes binary files, software libraries, and/or other
resources required to execute applications on a guest operating
system (e.g., within the virtual computing system 306) or the host
operating system 305.
[0060] Although the detection system 301 may include one or more
elements 303-313 of the cloud computing system 302, may execute
within the cloud computing system 302, and/or may be hosted within
the cloud computing system 302, in some implementations, the
detection system 301 may not be cloud-based (e.g., may be
implemented outside of a cloud computing system) or may be
partially cloud-based. For example, the detection system 301 may
include one or more devices that are not part of the cloud
computing system 302, such as device 400 of FIG. 4, which may
include a standalone server or another type of computing device.
The detection system 301 may perform one or more operations and/or
processes described in more detail elsewhere herein.
[0061] Network 320 includes one or more wired and/or wireless
networks. For example, network 320 may include a cellular network,
a public land mobile network (PLMN), a local area network (LAN), a
wide area network (WAN), a private network, the Internet, and/or
the like, and/or a combination of these or other types of networks.
The network 320 enables communication among the devices of
environment 300.
[0062] User device 330 includes one or more devices capable of
receiving, generating, storing, processing, and/or providing
information, as described elsewhere herein. User device 330 may
include a communication device and/or a computing device. For
example, user device 330 may include a wireless communication
device, a user equipment (UE), a mobile phone (e.g., a smart phone
or a cell phone, among other examples), a laptop computer, a tablet
computer, a handheld computer, a desktop computer, a gaming device,
a wearable communication device (e.g., a smart wristwatch or a pair
of smart eyeglasses, among other examples), an Internet of Things
(IoT) device, or a similar type of device. User device 330 may
communicate with one or more other devices of environment 300, as
described elsewhere herein.
[0063] The number and arrangement of devices and networks shown in
FIG. 3 are provided as an example. In practice, there may be
additional devices and/or networks, fewer devices and/or networks,
different devices and/or networks, or differently arranged devices
and/or networks than those shown in FIG. 3. Furthermore, two or
more devices shown in FIG. 3 may be implemented within a single
device, or a single device shown in FIG. 3 may be implemented as
multiple, distributed devices. Additionally, or alternatively, a
set of devices (e.g., one or more devices) of environment 300 may
perform one or more functions described as being performed by
another set of devices of environment 300.
[0064] FIG. 4 is a diagram of example components of a device 400,
which may correspond to detection system 301 and/or user device
330. In some implementations, detection system 301 and/or user
device 330 may include one or more devices 400 and/or one or more
components of device 400. As shown in FIG. 4, device 400 may
include a bus 410, a processor 420, a memory 430, a storage
component 440, an input component 450, an output component 460, and
a communication component 470.
[0065] Bus 410 includes a component that enables wired and/or
wireless communication among the components of device 400.
Processor 420 includes a central processing unit, a graphics
processing unit, a microprocessor, a controller, a microcontroller,
a digital signal processor, a field-programmable gate array, an
application-specific integrated circuit, and/or another type of
processing component. Processor 420 is implemented in hardware,
firmware, or a combination of hardware and software. In some
implementations, processor 420 includes one or more processors
capable of being programmed to perform a function. Memory 430
includes a random-access memory, a read only memory, and/or another
type of memory (e.g., a flash memory, a magnetic memory, and/or an
optical memory).
[0066] Storage component 440 stores information and/or software
related to the operation of device 400. For example, storage
component 440 may include a hard disk drive, a magnetic disk drive,
an optical disk drive, a solid-state disk drive, a compact disc, a
digital versatile disc, and/or another type of non-transitory
computer-readable medium. Input component 450 enables device 400 to
receive input, such as user input and/or sensed inputs. For
example, input component 450 may include a touch screen, a
keyboard, a keypad, a mouse, a button, a microphone, a switch, a
sensor, a global positioning system component, an accelerometer, a
gyroscope, an actuator, and/or the like. Output component 460
enables device 400 to provide output, such as via a display, a
speaker, and/or one or more light-emitting diodes. Communication
component 470 enables device 400 to communicate with other devices,
such as via a wired connection and/or a wireless connection. For
example, communication component 470 may include a receiver, a
transmitter, a transceiver, a modem, a network interface card, an
antenna, and/or the like.
[0067] Device 400 may perform one or more processes described
herein. For example, a non-transitory computer-readable medium
(e.g., memory 430 and/or storage component 440) may store a set of
instructions (e.g., one or more instructions, code, software code,
program code, and/or the like) for execution by processor 420.
Processor 420 may execute the set of instructions to perform one or
more processes described herein. In some implementations, execution
of the set of instructions, by one or more processors 420, causes
the one or more processors 420 and/or the device 400 to perform one
or more processes described herein. In some implementations,
hardwired circuitry may be used instead of or in combination with
the instructions to perform one or more processes described herein.
Thus, implementations described herein are not limited to any
specific combination of hardware circuitry and software.
[0068] The number and arrangement of components shown in FIG. 4 are
provided as an example. Device 400 may include additional
components, fewer components, different components, or differently
arranged components than those shown in FIG. 4. Additionally, or
alternatively, a set of components (e.g., one or more components)
of device 400 may perform one or more functions described as being
performed by another set of components of device 400.
[0069] FIG. 5 is a flowchart of an example process 500 for
utilizing multiple stacked machine learning models to detect
deepfake content. In some implementations, one or more process
blocks of FIG. 5 may be performed by a device (e.g., detection
system 301). In some implementations, one or more process blocks of
FIG. 5 may be performed by another device or a group of devices
separate from or including the device, such as a user device (e.g.,
user device 330). Additionally, or alternatively, one or more
process blocks of FIG. 5 may be performed by one or more components
of device 400, such as processor 420, memory 430, storage component
440, input component 450, output component 460, and/or
communication component 470.
[0070] As shown in FIG. 5, process 500 may include receiving
training data that includes input content for training machine
learning models that include a first machine learning model, a
second machine learning model, and a third machine learning model
(block 510). For example, the device may receive training data that
includes input content for training machine learning models that
include a first machine learning model, a second machine learning
model, and a third machine learning model, as described above.
[0071] As further shown in FIG. 5, process 500 may include
preprocessing the training data to generate preprocessed training
data (block 520). For example, the device may preprocess the
training data to generate preprocessed training data, as described
above.
[0072] As further shown in FIG. 5, process 500 may include training
the machine learning models, with the preprocessed training data,
to generate trained machine learning models that include a first
trained machine learning model, a second trained machine learning
model, and a third trained machine learning model (block 530). For
example, the device may train the machine learning models, with the
preprocessed training data, to generate trained machine learning
models that include a first trained machine learning model, a
second trained machine learning model, and a third trained machine
learning model, as described above.
[0073] As further shown in FIG. 5, process 500 may include
receiving particular input content (block 540). For example, the
device may receive particular input content, as described
above.
[0074] As further shown in FIG. 5, process 500 may include
processing the particular input content, with the first trained
machine learning model, to generate a first model result (block
550). For example, the device may process the particular input
content, with the first trained machine learning model, to generate
a first model result, as described above.
[0075] As further shown in FIG. 5, process 500 may include
processing the particular input content, with the second trained
machine learning model, to generate a second model result (block
560). For example, the device may process the particular input
content, with the second trained machine learning model, to
generate a second model result, as described above.
[0076] As further shown in FIG. 5, process 500 may include
processing the particular input content, with the third trained
machine learning model, to generate a third model result (block
570). For example, the device may process the particular input
content, with the third trained machine learning model, to generate
a third model result, as described above.
[0077] As further shown in FIG. 5, process 500 may include
processing the first model result, the second model result, and the
third model result, with an aggregation model, to generate a
detection result indicating whether the particular input content is
a deepfake or is real (block 580). For example, the device may
process the first model result, the second model result, and the
third model result, with an aggregation model, to generate a
detection result indicating whether the particular input content is
a deepfake or is real, as described above.
[0078] As further shown in FIG. 5, process 500 may include
performing one or more actions based on the detection result (block
590). For example, the device may perform one or more actions based
on the detection result, as described above.
[0079] Process 500 may include additional implementations, such as
any single implementation or any combination of implementations
described below and/or in connection with one or more other
processes described elsewhere herein.
[0080] In a first implementation, preprocessing the training data
to generate the preprocessed training data includes extracting
frames from the input content; extracting images of faces from the
frames; returning facial landmarks, associated with the images of
the faces, to the frames; and adding perturbations to the frames,
after returning the facial landmarks, to generate the preprocessed
training data.
[0081] In a second implementation, alone or in combination with the
first implementation, adding the perturbations to the frames
includes cropping outer pixels of the images of the faces and
replacing the cropped outer pixels with a border or resizing the
images of the faces based on the cropped outer pixels.
[0082] In a third implementation, alone or in combination with one
or more of the first and second implementations, each of the first
model result, the second model result, and the third model result
provides an indication of whether the particular input content is a
deepfake or is real.
[0083] In a fourth implementation, alone or in combination with one
or more of the first through third implementations, the first
machine learning model is a full-face image convolutional neural
network (CNN) model, the second machine learning model is a partial
image CNN model, and the third machine learning model is a support
vector machine learning model.
[0084] In a fifth implementation, alone or in combination with one
or more of the first through fourth implementations, the
aggregation model is a logistic regression machine learning
model.
[0085] In a sixth implementation, alone or in combination with one
or more of the first through fifth implementations, the first model
result is a first probability that the particular input content is
a deepfake and the second model result is a second probability that
the particular input content is a deepfake.
[0086] In a seventh implementation, alone or in combination with
one or more of the first through sixth implementations, processing
the input content, with the third machine learning model, to
generate the third model result includes processing the input
content, with the third machine learning model, to generate a
plurality of binary predictions that the input content is a
deepfake, and averaging the plurality of binary predictions to
determine a probability that the input content is a deepfake,
wherein the third model result corresponds to the probability.
[0087] In an eighth implementation, alone or in combination with
one or more of the first through seventh implementations,
performing the one or more actions includes one or more of
providing the detection result for display or retraining one or
more of the first machine learning model, the second machine
learning model, or the third machine learning model based on the
detection result.
[0088] In a ninth implementation, alone or in combination with one
or more of the first through eighth implementations, performing the
one or more actions includes one or more of generating and
providing a notification to a law enforcement agency when the
detection result indicates that the input content is a deepfake or
generating and providing a notification to an entity associated
with the input content when the detection result indicates that the
input content is a deepfake.
[0089] In a tenth implementation, alone or in combination with one
or more of the first through ninth implementations, performing the
one or more actions includes correcting the input content to remove
deepfake content and to generate modified input content and
utilizing the modified input content.
[0090] In an eleventh implementation, alone or in combination with
one or more of the first through tenth implementations, performing
the one or more actions includes preventing a financial transaction
when the detection result indicates that the input content is a
deepfake of a credential required for performing the financial
transaction.
[0091] In a twelfth implementation, alone or in combination with
one or more of the first through eleventh implementations, the
detection result is a binary result with a first value indicating
that the input content is a deepfake and a second value indicating
that the input content is real.
[0092] Although FIG. 5 shows example blocks of process 500, in some
implementations, process 500 may include additional blocks, fewer
blocks, different blocks, or differently arranged blocks than those
depicted in FIG. 5. Additionally, or alternatively, two or more of
the blocks of process 500 may be performed in parallel.
[0093] The foregoing disclosure provides illustration and
description but is not intended to be exhaustive or to limit the
implementations to the precise form disclosed. Modifications may be
made in light of the above disclosure or may be acquired from
practice of the implementations.
[0094] As used herein, the term "component" is intended to be
broadly construed as hardware, firmware, or a combination of
hardware and software. It will be apparent that systems and/or
methods described herein may be implemented in different forms of
hardware, firmware, and/or a combination of hardware and software.
The actual specialized control hardware or software code used to
implement these systems and/or methods is not limiting of the
implementations. Thus, the operation and behavior of the systems
and/or methods are described herein without reference to specific
software code--it being understood that software and hardware can
be used to implement the systems and/or methods based on the
description herein.
[0095] As used herein, satisfying a threshold may, depending on the
context, refer to a value being greater than the threshold, greater
than or equal to the threshold, less than the threshold, less than
or equal to the threshold, equal to the threshold, and/or the like,
depending on the context.
[0096] Although particular combinations of features are recited in
the claims and/or disclosed in the specification, these
combinations are not intended to limit the disclosure of various
implementations. In fact, many of these features may be combined in
ways not specifically recited in the claims and/or disclosed in the
specification. Although each dependent claim listed below may
directly depend on only one claim, the disclosure of various
implementations includes each dependent claim in combination with
every other claim in the claim set.
[0097] No element, act, or instruction used herein should be
construed as critical or essential unless explicitly described as
such. Also, as used herein, the articles "a" and "an" are intended
to include one or more items and may be used interchangeably with
"one or more." Further, as used herein, the article "the" is
intended to include one or more items referenced in connection with
the article "the" and may be used interchangeably with "the one or
more." Furthermore, as used herein, the term "set" is intended to
include one or more items (e.g., related items, unrelated items, a
combination of related and unrelated items, and/or the like), and
may be used interchangeably with "one or more." Where only one item
is intended, the phrase "only one" or similar language is used.
Also, as used herein, the terms "has," "have," "having," or the
like are intended to be open-ended terms. Further, the phrase
"based on" is intended to mean "based, at least in part, on" unless
explicitly stated otherwise. Also, as used herein, the term "or" is
intended to be inclusive when used in a series and may be used
interchangeably with "and/or," unless explicitly stated otherwise
(e.g., if used in combination with "either" or "only one of").
[0098] In the preceding specification, various example embodiments
have been described with reference to the accompanying drawings. It
will, however, be evident that various modifications and changes
may be made thereto, and additional embodiments may be implemented,
without departing from the broader scope of the invention as set
forth in the claims that follow. The specification and drawings are
accordingly to be regarded in an illustrative rather than
restrictive sense.
* * * * *